2 * Copyright (c) 2018 Sergey Lavrushkin
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * Filter implementing image super-resolution using deep convolutional networks.
24 * https://arxiv.org/abs/1501.00092
25 * https://arxiv.org/abs/1609.05158
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
33 #include "libavformat/avio.h"
34 #include "libswscale/swscale.h"
35 #include "dnn_interface.h"
37 typedef struct SRContext {
41 DNNBackendType backend_type;
42 DNNModule *dnn_module;
47 struct SwsContext *sws_contexts[3];
48 int sws_slice_h, sws_input_linesize, sws_output_linesize;
51 #define OFFSET(x) offsetof(SRContext, x)
52 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
53 static const AVOption sr_options[] = {
54 { "dnn_backend", "DNN backend used for model execution", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS, "backend" },
55 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
56 #if (CONFIG_LIBTENSORFLOW == 1)
57 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
59 { "scale_factor", "scale factor for SRCNN model", OFFSET(scale_factor), AV_OPT_TYPE_INT, { .i64 = 2 }, 2, 4, FLAGS },
60 { "model", "path to model file specifying network architecture and its parameters", OFFSET(model_filename), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
64 AVFILTER_DEFINE_CLASS(sr);
66 static av_cold int init(AVFilterContext *context)
68 SRContext *sr_context = context->priv;
70 sr_context->dnn_module = ff_get_dnn_module(sr_context->backend_type);
71 if (!sr_context->dnn_module){
72 av_log(context, AV_LOG_ERROR, "could not create DNN module for requested backend\n");
73 return AVERROR(ENOMEM);
75 if (!sr_context->model_filename){
76 av_log(context, AV_LOG_ERROR, "model file for network was not specified\n");
79 if (!sr_context->dnn_module->load_model) {
80 av_log(context, AV_LOG_ERROR, "load_model for network was not specified\n");
83 sr_context->model = (sr_context->dnn_module->load_model)(sr_context->model_filename);
86 if (!sr_context->model){
87 av_log(context, AV_LOG_ERROR, "could not load DNN model\n");
91 sr_context->input.dt = DNN_FLOAT;
92 sr_context->sws_contexts[0] = NULL;
93 sr_context->sws_contexts[1] = NULL;
94 sr_context->sws_contexts[2] = NULL;
99 static int query_formats(AVFilterContext *context)
101 const enum AVPixelFormat pixel_formats[] = {AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
102 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
104 AVFilterFormats *formats_list;
106 formats_list = ff_make_format_list(pixel_formats);
108 av_log(context, AV_LOG_ERROR, "could not create formats list\n");
109 return AVERROR(ENOMEM);
112 return ff_set_common_formats(context, formats_list);
115 static int config_props(AVFilterLink *inlink)
117 AVFilterContext *context = inlink->dst;
118 SRContext *sr_context = context->priv;
119 AVFilterLink *outlink = context->outputs[0];
120 DNNReturnType result;
121 int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
122 const char *model_output_name = "y";
124 sr_context->input.width = inlink->w * sr_context->scale_factor;
125 sr_context->input.height = inlink->h * sr_context->scale_factor;
126 sr_context->input.channels = 1;
128 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
129 if (result != DNN_SUCCESS){
130 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
134 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
135 if (result != DNN_SUCCESS){
136 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
140 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
141 sr_context->input.width = inlink->w;
142 sr_context->input.height = inlink->h;
143 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
144 if (result != DNN_SUCCESS){
145 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
148 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
149 if (result != DNN_SUCCESS){
150 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
153 sr_context->scale_factor = 0;
155 outlink->h = sr_context->output.height;
156 outlink->w = sr_context->output.width;
157 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
158 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
159 0, NULL, NULL, NULL);
160 sr_context->sws_input_linesize = sr_context->input.width << 2;
161 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
162 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
163 0, NULL, NULL, NULL);
164 sr_context->sws_output_linesize = sr_context->output.width << 2;
165 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
166 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
167 return AVERROR(ENOMEM);
169 if (sr_context->scale_factor){
170 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
171 outlink->w, outlink->h, outlink->format,
172 SWS_BICUBIC, NULL, NULL, NULL);
173 if (!sr_context->sws_contexts[0]){
174 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
175 return AVERROR(ENOMEM);
177 sr_context->sws_slice_h = inlink->h;
179 if (inlink->format != AV_PIX_FMT_GRAY8){
180 sws_src_h = sr_context->input.height;
181 sws_src_w = sr_context->input.width;
182 sws_dst_h = sr_context->output.height;
183 sws_dst_w = sr_context->output.width;
185 switch (inlink->format){
186 case AV_PIX_FMT_YUV420P:
187 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 1);
188 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
189 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 1);
190 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
192 case AV_PIX_FMT_YUV422P:
193 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
194 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
196 case AV_PIX_FMT_YUV444P:
198 case AV_PIX_FMT_YUV410P:
199 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 2);
200 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
201 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 2);
202 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
204 case AV_PIX_FMT_YUV411P:
205 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
206 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
209 av_log(context, AV_LOG_ERROR,
210 "could not create SwsContext for scaling for given input pixel format: %s\n",
211 av_get_pix_fmt_name(inlink->format));
214 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
215 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
216 SWS_BICUBIC, NULL, NULL, NULL);
217 if (!sr_context->sws_contexts[0]){
218 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
219 return AVERROR(ENOMEM);
221 sr_context->sws_slice_h = sws_src_h;
228 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
230 AVFilterContext *context = inlink->dst;
231 SRContext *sr_context = context->priv;
232 AVFilterLink *outlink = context->outputs[0];
233 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
234 DNNReturnType dnn_result;
237 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
239 return AVERROR(ENOMEM);
241 av_frame_copy_props(out, in);
242 out->height = sr_context->output.height;
243 out->width = sr_context->output.width;
244 if (sr_context->scale_factor){
245 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
246 0, sr_context->sws_slice_h, out->data, out->linesize);
248 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
249 0, out->height, (uint8_t * const*)(&sr_context->input.data),
250 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
252 if (sr_context->sws_contexts[0]){
253 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
254 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
255 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
256 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
259 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
260 0, in->height, (uint8_t * const*)(&sr_context->input.data),
261 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
265 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
266 if (dnn_result != DNN_SUCCESS){
267 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
271 sws_scale(sr_context->sws_contexts[2], (const uint8_t *[4]){(const uint8_t *)sr_context->output.data, 0, 0, 0},
272 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
273 0, out->height, (uint8_t * const*)out->data, out->linesize);
275 return ff_filter_frame(outlink, out);
278 static av_cold void uninit(AVFilterContext *context)
281 SRContext *sr_context = context->priv;
283 if (sr_context->dnn_module){
284 (sr_context->dnn_module->free_model)(&sr_context->model);
285 av_freep(&sr_context->dnn_module);
288 for (i = 0; i < 3; ++i){
289 sws_freeContext(sr_context->sws_contexts[i]);
293 static const AVFilterPad sr_inputs[] = {
296 .type = AVMEDIA_TYPE_VIDEO,
297 .config_props = config_props,
298 .filter_frame = filter_frame,
303 static const AVFilterPad sr_outputs[] = {
306 .type = AVMEDIA_TYPE_VIDEO,
311 AVFilter ff_vf_sr = {
313 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
314 .priv_size = sizeof(SRContext),
317 .query_formats = query_formats,
319 .outputs = sr_outputs,
320 .priv_class = &sr_class,
321 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC,